Tendon / Cable Drives

Tendon drives cross over in 2037, and they get there in a way no other component on this board does: not because the supply is constrained, but because the parts wear out. This is the most elastic supply chain in the report, the raw fibre is abundant to the point of irrelevance, the machinery is commodity, and the brownfield lead time is the shortest of any link. And it crosses anyway. The reason is that a tendon is a fatigue element, reported at 1 to 5 million cycles before replacement², so a fleet does not buy tendons once, it re-buys them for as long as it operates. By 2040 replacements are already 48 percent of cumulative demand. A humanoid carries roughly thirty terminated assemblies, anchored on a published 20-degree-of-freedom hand running 30 Bowden cables in 15 antagonistic pairs¹, and across the fleet ramp⁹ that compounds to about 2.7 billion assemblies of cumulative need by 2040, a 2.3× gap at the 2040 horizon with 44 percent of the 2040 need met. There is one genuine inversion here worth flagging early: this is the only early binder where the United States is a primary producer rather than a bystander, at both ends of the chain, which is why the US-only line crosses in 2032 rather than the 2025 to 2029 that every other binder shows. Two caveats govern everything below. The replacement clock, not the per-robot count, decides this result, and no published source sizes this component at all, in units or in dollars.

A tendon drive is the alternative answer to the same question a miniature ball screw answers: how do you move a finger? The screw approach puts a motor and a small screw inside or beside each finger. The tendon approach does the opposite, it remotes the actuator mass out of the hand entirely, into the forearm, and pulls the distal joints with braided cables routed through low-friction sleeves. That is what buys a tendon-driven hand its dexterity-to-weight ratio, and it is why the architecture persists despite being harder to build. The cables themselves are braided ultra-high-molecular-weight polyethylene, Dyneema and Spectra are the trade names, typically 0.8 to 2.0mm in diameter, achievable down to 0.3mm, carrying up to 2,500N at under 1.5 grams per metre². UHMWPE wins here on a specific technical point rather than on cost: braided fibre survives the 3 to 10mm pulley radii a humanoid hand forces on it, where steel cable of equivalent strength would fatigue and fail. Some designs still run steel-in-PTFE Bowden assemblies¹, and this link counts both. What it does not count is electrical harness, that is a separate component in this report, and the published “robot cables” market studies measure that, not this. The two architectures are also not strictly rivals: the stated direction for next-generation dexterous hands is a composite “tendon rope plus miniature ball screw” transmission⁸, using both.

Every other mechanical component in this report is a durable: fit it once, and cumulative need is simply the fleet multiplied by a per-robot count. Tendons are not, and that single difference drives the entire result. Cumulative need here is initial fitment plus a compounding replacement flow, modelled at a four-year mean service life. Integrating the consensus ramp⁹, fitment alone reaches about 1.4 billion assemblies by 2040, but replacements add a further 1.3 billion on top, so total need runs to about 2.7 billion by 2040. The replacement term is closing on fitment as the dominant driver by the end of the horizon. The four-year figure is the most consequential assumption in this section, and it is deliberately not tuned. It is taken from the tactile sensor link, the other hand-resident wear item this report models, precisely so the number is inherited rather than fitted. The published fatigue band gives the honest spread: at roughly 20,000 tendon cycles a day, warehouse-duty hands reach the low end of the 1 to 5 million cycle range² inside about a year, while a lightly-used fleet comfortably exceeds a decade. Halve the life and this link becomes one of the most severe on the board; double it and the crossover slides well past 2040. The per-robot count of thirty carries its own caveat: it blends a published tendon-driven hand at 30 cables¹ against an architecture-share estimate, because roughly half the fleet's hands are assumed to use miniature screw drives instead. That half is now a single shared parameter read by both this link and ball screws, when the two were first published it was not, and each carried its own silently incompatible assumption. It remains a modelling judgment rather than a measurement.

Here the geography inverts relative to every other component in this report. The high-end raw fibre is American. Avient's Dyneema line at roughly 14,200 tonnes a year and Honeywell's Spectra at roughly 3,200 tonnes are the two largest non-Chinese producers, ahead of Toyobo's 3,000 tonnes⁴, and the precision assembly tier is US-located too: Carl Stahl Sava in Riverdale NJ builds mechanical cable down to 0.006 inches for surgical robotics³, Loos & Co runs a 220,000 square foot plant in Pomfret CT⁶, alongside Bergen Cable in Fairfield NJ⁷ and Lexco in Chicago. China's position is the mirror image of its usual one: it holds the volume, about 67 percent of installed fibre capacity in 2023, built out to roughly 730,000 tonnes a year of nameplate by end-2025 at only 23 percent utilization, but under 3 percent of the high-end market⁵. The result is a US production share of about 20 percent, the highest of any link on the board. But US-located is not US-owned, and this is the finding that matters for anyone trying to invest in it. All three of the major American mechanical-cable houses sit inside larger parents: Sava is the US arm of Germany's Carl Stahl, Loos is the Rope and Assemblies division of Canada's Central Wire⁶, and Bergen was acquired by Leggett & Platt in June 2020⁷. The domestic capability is real; the domestic ownership is almost entirely absent. On the supply side there is no scarcity to speak of, global output runs from roughly 42 million assemblies a year today toward 115 million by 2040, and could go much faster if asked.

The durable crossover lands in 2037 globally, the latest of any binding component except the encoder and the two motor lines, which places tendons at the mild end of the binding set. The shortfall then deepens hard, precisely because the replacement flow compounds: 44 percent of the 2040 need met, with cumulative production needing to scale about 2.3× to close the 2040 gap. That shape, a late crossover followed by an accelerating rather than settling shortfall, is the signature of a consumable. The single-bloc views tell the story this component is really about. China-only crosses in 2034 and meets 16 percent of the 2040 need. US-only crosses in 2032, and that number deserves attention, because every other early binder crosses in the US between 2025 and 2029. On tendons the United States is five to seven years better off than on anything else it needs to build a robot hand. It is the one place in this report where the reshoring question is not “can America make this at all” but “can America own the companies that already do”.

It is worth being explicit about what is not constrained here, because the instinct with a shortage is to look for a scarce machine and there is none. The fibre is not scarce. Global UHMWPE fibre demand runs 70,000 to 80,000 tonnes a year⁴, while a humanoid's thirty tendons weigh a matter of grams, even a 47-million-robot fleet consumes on the order of 1,000 tonnes a year, well under two percent of a market that is itself sitting on enormous idle capacity: China alone built to 730,000 tonnes of nameplate at 23 percent utilization⁵. Robots will never move the fibre market. The machinery is not scarce either. A braider is an 8- to 24-spindle machine running 0.1 to 5 metres per minute², commodity capital from many vendors, not a long-lead precision tool, and suppliers have already automated cut-to-length, swaging and fitting specifically to scale volume³. This link therefore carries the shortest brownfield lead time and the lowest capital intensity on the board, and it obviously has nothing to do with the shared precision-grinding pool that throttles roller screws, harmonic drives and cross-roller bearings; a braided textile part never touches a thread grinder. What is actually constrained is engineering, not tonnage. UHMWPE resists conventional crimping, so terminations need braided or woven sleeves with epoxy potting or purpose-designed mechanical interlocks, and holding fatigue life requires pulley-to-cable diameter ratios of at least 25׳, a specification a compact humanoid hand fights constantly. The scarce thing is the know-how that certifies a million-cycle life at sub-millimetre diameters. That is a qualification problem, and qualification problems are solved by engineers rather than by capital expenditure, which is why this component is graded elastic, and why, if the wear clock proves longer than four years, this crossover is the most likely in the report to simply disappear.